Accurate and Low-cost Indoor Location Estimation Using Kernels - - PowerPoint PPT Presentation

1

Accurate and Low-cost Indoor Location Estimation Using Kernels

Jeffrey J. Pan, James T. Kwok, Qiang Yang, Yiqiang Chen Department of Computer Science Hong Kong University of Science and Technology

Present in the Nineteenth International Joint Conference on Artificial Intelligence (IJCAI-05), Edinburgh, Scotland, July 2005.

2

HKUST

Application Background

Positioning

• Outdoor : Road Guiding (GPS)
• Indoor : Large Building (WiFi)

Location-based Service

• Web Content Delivery

Behavior Analysis

• Daily Life (L. Liao et al. AAAI-04, IJCAI-05)
• Health Care
• Scientific Purpose

3

HKUST

Problem Description

• A user with a mobile device walks in an indoor

wireless environment (Covered by WiFi signals)

AP1

AP2 AP3 Where am I ? Time t: (-47dB,-36dB,-62dB) Off-the-Shelf Hardware, Only Signal Strength

4

HKUST

Noisy Propagation Channel at 2.4G

AP1 AP2 AP3

Signal Obstructed Signal Obstructed

Radio Map

5

HKUST

Learning-based Location Estimation

• Two phases: offline Training and online Localization
• Offline phase – collect samples to build a mapping function F

from signal space S to location space L

• Online phase – given a new signal s , estimate the most likely

location l from F

• s = (-60,-49,-36)dB , compute F(s) as the estimated location

(-50,-35,-42) dB 9s (9,5) ( … , … , … )dB …. ….. (-62,-48,-35) dB 2s (2,0) (-60,-50,-40) dB 1s (1,0) (AP1,AP2,AP3) Time Loc.

Mapping function F Training…

6

HKUST

Outline

• Introduction to Location Estimation
• Application Background
• Problem Description
• Noisy Characteristics of Propagation Channel
• Basic Framework for Location Estimation
• Related Work
• Microsoft Research’s RADAR (INFOCOM’2000)
• University of Maryland’s Horus (PerCom’2003)
• Motivation of Our Approach
• The LE-KCCA Algorithm
• Kernel Canonical Correlation Analysis (KCCA)
• Choices of Kernels
• Experimental Setup and Result
• Strength and Weakness
• Future Work

7

HKUST

Related Works

• Microsoft Research’s RADAR [P. Bahl et al. INFOCOM2000]
• K-Nearest-Neighbor Method
• Offline - for each location, compute the signal mean
• Online – estimate location with KNN and triangulation

Strength

• Small number of samples could estimate the signal mean well

Weakness

• Accuracy is relatively low
• Reason – The K nearest neighbors retrieved in the signal space may

not necessarily the K nearest neighbors in the location space

8

HKUST

Related Works (Cont’)

• University of Maryland’s Horus [M. Youssef et al. ,2003]
• Maximum Likelihood Estimation (MLE)
• Offline - for each location, build the Radio Map of each AP
• Online - apply Bayes’ rule for estimation
• Strength
• Accuracy is high
• Weakness
• Need relatively large number of samples
• Reason – More samples are needed for establishing an accurate

Radio Map rather than a signal mean

Radio Map

9

HKUST

Motivation of Our Approach

Observation (Motivated by RADAR)

Similar signals may not be nearby locations Dissimilar signals may not be far away

Idea

Maximize the similarity correlation between signal

and location spaces under feature transformation

Goal

Accuracy as high as possible ( Horus ) Calibration Effort as low as possible ( RADAR )

10

HKUST

Motivation of Our Approach (Cont’)

Original Signal Space Feature Signal Space Original Location Space Feature Location Space

11

HKUST

(Kernel) CCA

• Canonical Correlation Analysis (CCA)
• [H. Hotelling, 1936]
• Two data set X and Y
• Two linear Canonical Vectors Wx Wy
• Maximize the correlation of projections
• Kernel CCA
• [D.R Hardoon, S. Szedmak, and J.

Shawe-Taylor, 2004]

• Two non-linear Canonical Vectors
• K is the kernel

12

HKUST

Offline phase

• Signal strengths are collected at various grid locations.
• KCCA is used to learn the mapping between signal and location spaces.
• λi’s and αi’s are obtained from the generalized eigen-problem
• κis a regularization term
• For each training pair (si, li), its projections
• n the T canonical vectors are obtained from

LE-KCCA

13

HKUST

LE-KCCA (Cont’)

• Online phase
• Assume the location of a new signal strength vector is s
• Again, use

to project s onto the canonical vectors and obtain

• Find the K Nearest Neighbors of P(s) in the projections of training

set with the weighted Euclidean distance :

• Interpolate these neighbors’ locations to predict the location of s
• Essentially, we are performing Weighted KNN in the feature space with

which weights are obtained from the feedback of location information.

14

HKUST

Choices of Kernels

• Kernel for Signal Space
• Gaussian Kernel to smooth the noisy characteristics
• Widely used : [Roos et al. 2002, Battiti et al. 2002]
• Kernel for Location Space
• Matern Kernel to sense the change in location
• Used in : GPPS [Schwaighofer et al., 2003]

15

HKUST

Experimental Setup

• Test-bed : Department of Computer Science, Hong

Kong University of Science and Technology

99 locations (1.5 × 1.5 meter) 100 samples per location 65% for training, 35% testing Repeat each experiment 10 times

65m 42m

16

HKUST

Experimental Result - 1

• Data Set
• 65% training
• 35% testing
• Error Distance is 3.0m
• LE-KCCA 91.6%
• SVM 87.8%
• MLE 86.1%
• RADAR 78.8%

Accuracy

17

HKUST

Experimental Result - 2

Reduce Calibration Effort

• Incrementally Use a small subset of the the 65% training data
• Outperform the others using 10-15 samples from each location

18

HKUST

Recall our Motivation………

Original Signal Space

Feature Signal Space

Original Location Space Feature Location Space

We could see on the next page……

19

HKUST

Visualization of Tracking in Both Original and Feature Spaces

10 20 30 40 50 60 70 10 20 30 40 50 60 70 AP1 (unit:−dB) AP2 (unit:−dB) 10 20 30 40 50 5 10 15 20 25 30 35 40 X (unit:1.5m) Y (unit:1.5m) −15 −10 −5 5 10 15 20 25 −20 −15 −10 −5 5 10 15 Feature 1 Feature 2 −15 −10 −5 5 10 15 20 25 −20 −15 −10 −5 5 10 15 Feature 1 Feature 2

Hallway 1 Hallway 2 Hallway 3 Hallway 1 Hallway 1 Hallway 1 Hallway 2 Hallway 2 Hallway 2 Hallway 3 Hallway 3 Hallway 3

Original Signal Space Feature Signal Space Original Location Space Feature Location Space

20

HKUST

Strength and Weakness

Strength

Higher Accuracy Reduced Calibration Effort (Low-cost)

Weakness

Generally 50-100 times slower than RADAR

21

HKUST

Future Work

• Consider Environment Dynamics to Reduce Uncertainty
• J. Yin et al. Adaptive temporal radio maps for indoor location
• estimation. PerCom’05
• Consider User Dynamics to Reduce Uncertainty
• M. Berna et al. A Learning Algorithm for Localizing People Based On

Wireless Signal Strength That Uses Labeled and Unlabeled Data. IJCAI’03

• A. Ladd et al. Robotics-based location sensing using wireless ethernet,

MobiCom’02

• Speed up for Large-Scale Localization
• J. Letchner et al. Large Scale Localization from Wireless Signal
• Strength. AAAI’05
• A. Haeberlen et al. Practical Robust Localization over Large-Scale

802.11 Wireless Networks. MobiCom’04

22

HKUST

Acknowledge

Hong Kong RGC

HKUST6187/04E

Thanks Jie Yin and Xiaoyong Chai

Data collection Helpful discussion

23

HKUST

Thank You

Question ?